Efficient discovery of contrast subspaces for object explanation and characterization

Abstract

We tackle the novel problem of mining contrast subspaces. Given a set of multidimensional objects in two classes C + and C − and a query object o, we want to find the top-k subspaces that maximize the ratio of likelihood of o in C + against that in C − . Such subspaces are very useful for characterizing an object and explaining how it differs between two classes. We demonstrate that this problem has important applications, and, at the same time, is very challenging, being MAX SNP-hard. We present CSMiner, a mining method that uses kernel density estimation in conjunction with various pruning techniques. B Lei DuanWe experimentally investigate the performance of CSMiner on a range of data sets, evaluating its efficiency, effectiveness, and stability and demonstrating it is substantially faster than a baseline method.

Tree-based mining contrast subspace

“…Since it is not practical to manually examine a large number of features represents most of the real world data sets, mining contrast subspace has emerged to automate the process of discovering such abovementioned subspaces of an object. Given a multidimensional data set of two classes, a query object and a target class, mining contrast subspace finds subspaces where the query object is most similar to the target class while most dissimilar from other class [7], [8]. Those subspaces are also termed as contrast subspaces in the literature and it will be used throughout this paper.…”

“…There are only few methods have been developed to mine contrast subspaces of a query object. Traditional mining contrast subspace use a scoring function named likelihood contrast measure to quantify the likelihood contrast degree or likelihood contrast score of a subspace with respect to a given query object [7], [8]. The likelihood contrast score reflects to what extent a given query object similar to the target class against other class in a subspace.…”

“…The existing mining contrast subspace methods use density-based likelihood contrast measure [7], [8]. It defines contrast subspace as where a query object is situated in the area with greater density of target object (i.e.…”

“…CSMiner-BPR has been proposed to speed up the mining contrast subspace process of CSMiner [8]. It implements a number of bounding pruning strategies in addition to the existing pruning rule in CSMiner.…”

“…By knowing those subspaces, it helps to ensure right diagnosis in patient and appropriate treatments to be provided. Another example is in an insurance company, an analyst may want to know the subspaces that make a suspicious claim most similar to fraudulent cases but dissimilar to normal cases [7], [8]. Those subspaces give analyst useful information to perform further investigation.…”